Settings

Press ⌘, or choose Bandicot › Settings…; the window opens on the tab you used last.

Settings has twelve tabs in a row along the top: Input, Model, Agent, AI usage, Session, Appearance, Folders, Capabilities, Setup, Sign-in, Notifications and Your phone. Profiles and Memory are not tabs; they open from Home.

At the foot of each tab, Open usage shows or hides the usage panel beside the chat, the Context tab of the Inspector.

A control greyed out with Set by your environment (NAME=value); change it where Bandicot is started. is fixed by a BANDICOT_ variable in the shell that started Bandicot. Remove the variable and start Bandicot again to change it here.

The Input tab chooses what the prompt bar shows, how Enter behaves and which saved setups you can restore in one click.

Preview is the real prompt bar: its model and effort menus change the open session. Prompt bar holds the switches below. Show everything turns every part back on, and Just the field hides them all. The choices are kept for this profile.

A Tools card sits below the prompt bar choices: the same switches as on the Capabilities tab, so changing one here changes it there. Each row shows its tool name and how much that tool was used in the last thirty days, and Web and browser adds a count of browser actions.

Saved setups stores a model, an effort and the prompt bar layout under a name. Everything visible, Focus and Words only are built in and cannot be deleted. Type a name in Name this setup and press Save current setup to add yours, then Apply to restore it.

The Model tab picks the provider and model for the session and shows each provider's account status, with Connect and Disconnect.

Provider & model sets the session's provider and model. Accounts lists every provider with one status line, and a button or word at its right.

Disconnect signs you out of a provider so you can connect again with another account. Press it once and it turns into Cancel and Disconnect; press Disconnect again to confirm. The button reads Disconnecting… while it works, and a line under the list then reads Disconnected. Connect again to sign in with another account. If it fails, the line shows the reason, or Could not disconnect.

Connect a model walks through every way to connect.

The Agent tab sets how the agent answers and thinks, and holds your style, the project handbook and each profile's instructions.

Behavior offers four presets: Concise, Balanced (recommended and the default), Thorough and Cautious. Each sets four switches you can also change one by one. A mix that matches no preset reads Custom. Changes apply from the next chat; thinking depth applies from the next message.

Confirm risky actions is always on: deleting, overwriting, sending, publishing and paying go through the permission gate whatever the switches say.

Instructions has three texts. Your style follows you into every profile and project. This project's handbook is saved in the project's AGENTS.md, or in an existing CLAUDE.md, and wins over the other two. Instructions for this profile start every new chat in that profile. Each saves as you type and has Reset to recommended. New chats use changes at once; press Apply to this chat to send them to the open one.

Each of the three texts shows a badge: Recommended while it is unchanged, Edited once you change it, or Not set. On an edited text, Show changes (then Hide changes) compares it with the recommended text. If the project handbook sits below other AGENTS.md files higher in the repository, Also read, from higher in the repository: lists them, each with Open file.

Under Using them, a fold reads Final prompt · about N tokens, with a count in place of N. Open it to read exactly what the model is sent: the system prompt (core, mode and your style, marked Fixed, cached) and this chat's own block (workspace, profile and project), each with its own token estimate. The counts are estimates.

Models shows which settings the current model accepts, such as thinking depth or temperature, and which Bandicot manages for you.

The AI usage tab sets one dollar allowance for the AI work Bandicot starts by itself, counted at each model's API price.

Automatic work includes System One, setup, grading, experiments and reviews. On a signed-in provider account it only uses that account's own limits; real money is spent only when the model runs on an API key.

Type an amount in Automatic AI allowance, keep Enable automatic AI work within this allowance ticked and press Save account limit. The line below shows what is spent and reserved. The allowance does not reset by itself: raising it adds headroom. Usage explains the readouts.

The Session tab shows what the agent may do without asking and lets you edit permissions.json, the file that changes it.

The tab's card is Session defaults. Permissions reads Auto, the only mode: ordinary work runs without asking. Three rows show a verdict: Dangerous shell lines, Writes outside the project and Changes after reading team or web content. Each reads Runs, Asks you or Blocked.

Two more rows are information, not verdicts. Untrusted sources lists which tools count as untrusted (team and web unless your file changes it), and Project rules shows which permissions.json the project uses, by default <project>/.bandicot/permissions.json. Thinking depth, answer length and initiative are in the Agent tab.

The editor below holds your permissions.json. Edit it, then press Save or ⌘S. Approvals explains each rule and every message the editor can show.

Appearance switches the window between Dark, Light and Match macOS, and the window repaints at once.

Dark is the default: warm near-black with stone-white text. Match macOS follows the system setting as it changes through the day. The choice is kept per profile.

Folders lists folders the agent may read, search and, if you choose, write in without asking.

Press Add folder…, pick the folder, then set its access to read or write; new folders start as read. Remove takes a folder off the list. A change made while a turn runs applies after the turn.

The list is the same one the profile's config file holds, so editing either shows in both.

Capabilities turns each agent tool on or off and holds the switches for computer use, System One, compaction, memory lookup and sub-sessions.

Tools: a switched-off tool leaves the model's tool list, so its description stops using context on every turn. Every tool starts on except Diagrams and Canvas. The Versions group decides whether the agent may read or restore this project's history; you always can in Versions.

Computer use shows whether Accessibility and Screen capture are allowed, with buttons to request each. System One picks the decision model and connects TypeSafe with a key.

Context compaction is on by default; with it off, compaction never runs. With Selection compaction on, System One chooses what to drop instead of a model rewriting the history, and Context kept per compaction sets how much stays.

Brain efficiency holds Context lookup, Turn updates and Browser access, all on by default. Agent capabilities holds Reach outside the project (on), Continue down the plan chain (off) and Start sub-sessions without asking (off).

Decision Lab is experimental and off by default. In shadow mode it records what a decision model would answer and never acts on it.

A setup file describes your model, permissions, folders, skills, agents, memory, rules and hooks, and Setup shows every change before anything is written.

Press Choose a setup file…, or drop a setup.json, or the folder that holds it, onto the window. The review lists each change by group. Sensitive changes, such as hooks, allow rules, write folders and plugins, stay unticked until you tick them. Keys the file needs are typed into the review and stored only in Bandicot's secrets store, never in the file.

History lists every apply with Undo last apply. Export to folder… writes your skills, agents and memory into a setup.json for another Mac; keys are named, never copied. The same work runs from Terminal with bandicot setup.

If your hooks file changes outside Bandicot, a card reads Your hooks are paused. Read the file, then press the I’ve checked it button to turn hooks back on.

Bandicot also watches ~/.config/bandicot/setup.json while it is open, about every twenty seconds. When that file has changes waiting, a toast reads A setup file is ready to review; if you applied one before and it has changed, Your setup file changed. It shows how many changes are waiting and says nothing is applied until you choose. Review opens this tab. You get one toast for each distinct set of changes, and a paused hooks file raises Your hooks are paused with the same Review button.

1. 1Open Settings › Setup.
2. 2Press Choose a setup file… and pick the file.
3. 3Read the review. Tick each sensitive change you want and type any key it asks for.
4. 4Press Apply. A notice confirms the changes, with Undo.

The Sign-in tab opens your browser to set up a passkey for future sign-ins on this Mac.

Sign in with your email and password first, then press Set up or manage passkeys. The browser opens your account's passkey page on the website.

The Notifications tab has one control, Test banner, which sends a macOS banner so you can check Bandicot is allowed to notify you.

Press Test banner. A banner reading Notifications are working. should appear, and a line below confirms it was sent. If no banner appears, allow Bandicot in System Settings › Notifications. Bandicot's own notices also stay in the Activity list in the sidebar.

The button works only in the desktop app. In any other view the tab reads Native banners need the Bandicot desktop app. and the button is greyed out.

The Your phone tab keeps this Mac reachable from your phone and, on a laptop, working with the lid closed.

Turn on Reachable from my phone and keep the Mac plugged in. A checklist shows what else keeps it reachable, with a tick read from macOS for each. Your phone explains the lid-closed helper and its safety limits.

A setting shown as Set by your environment comes from a variable where Bandicot was started, and it wins over Settings until you remove it there.

The Agent tab lists each such variable under Set by your environment, and the control it sets stays read-only. The installed app does not read your shell profile, so a variable reaches it only when you start Bandicot from that shell.

A provider key such as OPENCODE_API_KEY or OPENAI_API_KEY in the environment shows as from your environment beside that provider in Sign-in.